Селекция объектов по отношениям с ближайшими объектами противоположного класса и сохранение локальной структуры обучающей выборки

The paper considers instance selection for a training sample and the analysis of how selection affects the local structure of the data. The ENRS (Enemy-Neighbor Relation Selection) algorithm is proposed, in which the decision to remove an object is based on its relations with the nearest objects of the opposite class and with same-class objects whose spheres of influence contain it. The local structure of a sample is described by a feature formed from the composition of the ???? nearest neighbours of an object and by the stability indicator of this feature; the set of indicator values for different ???? represents the sample as a meta-object in a unified space. A method for assessing the preservation of local structure is proposed, based on comparing a reduced sample with random subsamples of the same size and class composition. Computational experiments were carried out on the German Credit data set and five other UCI data sets, and ENRS was compared with the CNN, ENN and Tomek Links selection methods. It is found that on the German Credit and Pima data sets the samples produced by ENRS differ from random subsamples by a lower class homogeneity of object neighbourhoods, that the extent of this difference depends on the algorithm parameter, and that the classification accuracy on ENRS samples does not exceed that on random subsamples of the same size. The proposed approach separates changes of local structure caused by the selection of objects from those caused by the reduction of sample size and can be applied to the analysis of any selection method.